UniJev
This repository contains the UniJev model family introduced in UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models.
Code: https://github.com/wjnwjnwj/UniJev
Demo: https://unijev-website.3323422990.workers.dev/
Overview
UniJev is a trimodal decision framework that scores candidate answers with a shared head under joint supervised training across text, images, and video. Given a context (text, image, or video) and a set of candidate answers, the model returns an explicit probability distribution over the candidates, enabling applications to use the confidence score for routing, filtering, or requesting human review.
The framework consists of two training stages:
- Soft-label SFT — builds the candidate-scoring/decision capability.
- CORA (Confidence–Outcome Reliability Alignment) — a supervised calibration objective that groups predictions by confidence and optimizes each prediction's confidence toward the empirical accuracy of its bin, without a reinforcement learning stage.
Model sizes
| Backbone | Modalities | Parameters |
|---|---|---|
| Qwen3.5-2B | Text, image, video | 2B |
| Qwen3.5-9B | Text, image, video | 9B |
Both variants use Qwen3.5 backbones with LoRA adaptation, a trainable vision merger, and a lightweight decision head.
Training data
Training uses a shared JSONL schema for all three modalities. Each record contains a context, one or more questions with candidate answers (choice, binary judgment noul, or ordinal score), and target probabilities.
Citation
@misc{wang2026unijev,
author = {Wang, Jianing and Jian, Youwen and Wang, Derong and Yu, Jinyang and Si, Qingyi and Qin, Chuanyu and Xiao, Yanghua and Guo, Hongcheng},
title = {UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models},
year = {2026},
howpublished = {\url{https://github.com/wjnwjnwj/UniJev}},
note = {Research manuscript}
}